My college essay has been flagged as AI. What should I do?

Why AI Detectors Flag Honest Students

A flag on your essay feels like a verdict, but it isn't one. An AI detector score is a probabilistic estimate, not proof — and the institutions that know these tools best are increasingly refusing to trust them.

Consider the track record. Vanderbilt University disabled Turnitin's AI detector in August 2023, stating plainly that it does not believe AI detection software is effective — after calculating that even a 1% false positive rate would have wrongly flagged roughly 750 of the 75,000 papers its students submitted in a single year. OpenAI shut down its own detector in July 2023 over its "low rate of accuracy," and more than 25 universities, including MIT, Yale, and UC Berkeley, have banned or restricted detection tools outright.

The numbers behind those decisions are sobering. Vendors claim false positive rates under 1% — Turnitin's own whitepaper reports 0.51%, meaning about 1 in 200 honest submissions gets flagged. But independent testing tells a worse story, measuring rates of 15–26% against Turnitin's claimed 1%. And the errors aren't random — they cluster around specific types of students:

• Non-native English speakers: a Stanford study published in Patterns found detectors falsely flagged 61.3% of TOEFL essays written by Chinese students, versus just 5.2% for US-born writers.
• Neurodivergent students: a University of San Diego law library guide reports that students with autism, ADHD, or dyslexia are flagged at higher rates than their peers.
• Grammarly and translation-tool users: grammar and translation software can produce patterns detectors mistake for ChatGPT — one documented appeal involved a student flagged "100% AI" whose university-approved spell-checker was the culprit.

The tools also fail in ways that should embarrass their makers. AI detectors have flagged the US Constitution as AI-written. Turnitin offers no explanation of what patterns it looks for — Vanderbilt called it effectively a black box — and its own fine print says flagged text only "might be" AI-generated. As one Birkbeck lecturer put it after a student of his was wrongly accused, instructors making these calls are acting "without reliable evidence."

That's why the practical advice is shifting from "avoid detection" to "document your process." Students now create millions of authorship reports and keep draft histories before submission — and the same logic applies to any AI-assisted workflow. If you use AI tools for organizing drafts or tracking revisions, agent skills can automate that record-keeping; SkillzBench is building a free directory where you can browse and test such skills. The bottom line: if your honest essay got flagged, you're in good company, and the evidence is on your side.

Build Your Evidence File: What Wins Appeals

A flagged essay feels like a verdict, but it isn't one — it's a probability estimate that you can overturn with the right documentation. The most instructive case comes from Birkbeck, University of London, where a student flagged as "100% AI-generated" won their appeal on the strength of one thing: pre-existing evidence of their own writing style, established long before ChatGPT existed, as documented by their lecturer.

Start with your writing history, because it's the single most decisive factor in a successful appeal. Gather everything that shows the essay evolving in your own words:

• Early drafts and outlines, however rough or incomplete
• Google Docs version history showing incremental edits over time
• Prior assignments written in the same voice and style
• Notes, research clippings, or annotated sources behind your argument

Keeping drafts and version history is fast becoming standard practice — Grammarly reports over 5 million Authorship reports created by students in the past year, mostly for self-verification before submission.

Next, disclose any legitimate tools you used, proactively and in writing. In the Birkbeck case, the flag was traced to university-approved spelling and grammar software with limited generative AI capabilities that the detector mistook for ChatGPT. If you used translation or grammar support, say so up front — Stanford research shows these tools are a documented cause of false flags, especially for non-native English writers.

Finally, cite the bias research directly in your defense. The Stanford study found a 61.3% mean false positive rate for non-native English writers, versus 5.2% for US students, and detectors have even flagged the US Constitution as AI-written. Even Turnitin's own self-reported figure, per a vendor analysis, implies roughly 1 in 200 submissions is falsely flagged. Vanderbilt's official position is blunt: the university disabled Turnitin's detector, stating it does not believe AI detection software is an effective tool that should be used.

Courts are moving the same direction, establishing that detection scores alone don't constitute evidence of academic dishonesty. A score is an accusation, not proof — and a well-documented writing history is what turns the burden back where it belongs. If you want help automating version tracking and document workflows for future submissions, SkillzBench is building a free directory of agent skills you can browse and test for exactly that kind of task.

The Appeal Conversation: Step-by-Step

A flagged essay can feel like a verdict, but the institutions that have examined these tools most closely say a detector score is a starting point for conversation — not proof. Vanderbilt, which disabled Turnitin's AI detector after concluding it was ineffective, recommends that instructors respond to suspected AI use by talking with the student directly and comparing the flagged work against their previous writing, rather than issuing automatic punishment.

Step 1: Request a conversation, not a verdict. Email your instructor promptly — stakes in these cases can reach penalties as severe as expulsion, so treat the process with urgency and professionalism. Ask for a meeting to discuss the flag, and frame it as a chance to review evidence together. Vanderbilt's guidance specifically calls for comparing style, tone, and level against a student's prior work — so make that comparison easy to do.

Step 2: Bring your documentation. The single most decisive factor in a documented successful appeal was the student's pre-existing writing style from before ChatGPT existed. Gather:

• Drafts and version history (Google Docs revision history, Word tracked changes) showing the essay evolving over time
• Prior assignments written in the same voice, ideally from before AI tools were widespread
• A list of legitimate tools you used — spelling, grammar, or translation software that can trigger false flags
• Research notes, sources, and outlines that show your thinking process

Disclose any assistive tools proactively. In the Birkbeck case, the flag was traced to university-approved grammar software with limited generative AI features, and the student won on appeal. A Stanford study found detectors misclassified 61.3% of TOEFL essays by non-native English writers, so translation and grammar tools are a documented cause of false flags.

Step 3: Know the detector's limits — and your escalation options. Turnitin's own fine print says flagged text only "might be" AI-generated, and its AI report is a black box with no verifiable evidence, unlike a similarity report. If the informal conversation fails, request a formal appeal through your institution's academic integrity office or honor board, where you can present your documentation and the documented false-positive research.

Keep every draft going forward — over five million students created Grammarly Authorship reports last year to self-verify before submission, and version history is becoming standard self-defense. If you work with AI tools regularly, agent skills can automate document organization and version tracking; SkillzBench is building a free, verified directory of such skills, with security scanning and benchmarking, plus a free MCP connector so your agent only sees what you choose.

Protect Yourself Going Forward

Winning an appeal once is relief. Building a documentation habit means you never have to rely on panic and luck again — and the evidence is clear that process records are what tip these cases.

The documented successful appeal at Birkbeck turned on one thing: the student could show they'd written in the same style long before ChatGPT existed. That only works if the record already exists when the accusation lands. Start treating every essay as potential evidence.

Make these habits automatic:

• Keep version history enabled everywhere — Google Docs revision history is the single most persuasive artifact you can produce.
• Save drafts at every stage: outline, rough draft, edits, final. Gaps in the timeline are what prosecutors ask about.
• Document any tool you use — spellcheckers, Grammarly, translation apps. One appeal succeeded because a flag was traced to university-approved grammar software with limited generative features.
• Write in your established voice. Short, formulaic text is more prone to false flags, so your natural style is both your signature and your shield.

You don't have to manage this manually forever. Agent skills — instruction packages that extend tools like Claude Code, Codex, and Cursor — are emerging to automate document organization, version tracking, and writing workflows. SkillzBench is building a free, independent directory of these skills plus a free MCP connector so you can browse, verify, and test them; you can join the waitlist at skillzbench.com.

This kind of self-defense is becoming mainstream, not paranoid. Grammarly reports over 5 million Authorship reports created by students in the past year, most of them self-checks run before submission rather than responses to accusations. As one educator put it, students are now trying to prove they're human even when they've never touched AI.

That's the new baseline. Detection scores are probabilistic estimates, not findings of fact — but the student who can produce a draft trail never has to make that argument from scratch. Build the trail now, and the next flag becomes a five-minute conversation instead of a semester-long ordeal.